Extrapolation Ability of Denoising Diffusion Probabilistic Models for High-Frequency Inductor Core Loss Predictions With DC Bias

Authors:

Abstract
Accurate prediction of high-frequency magnetic core losses under DC bias conditions remains a critical challenge in power electronics design. Conventional models, such as empirical curve-fitting approaches and neural networks, often struggle with extrapolation beyond training data, particularly under varying DC bias levels. This paper proposes a Denoising Diffusion Probabilistic Model (DDPM) to enhance the extrapolation capability for core loss predictions in ferrite materials, N87 and 3C90, under DC bias. The proposed DDPM leverages a two-phase noise injection and denoising process to improve prediction stability and generalization. Experimental validation demonstrates that the proposed model achieves an R2 score of 0.9999 for interpolation and maintains high accuracy (R2 = 0.9996)in extrapolation tasks, significantly outperforming conventional CNN (R2 = 0.1856) and GANET (R2 = 0.7974) models. Additionally, the DDPM model enables high-fidelity data augmentation, generating realistic core loss datasets for unseen DC bias conditions. These results highlight the DDPM’s potential as a robust tool for magnetic component design, offering improved accuracy and reliability in both interpolation and extrapolation scenarios.
Funding Information
  • Research on Fault Location and Early Defect Warning Technology (HXKT2025-29-01)